Review

Definition of prognostic and predictive biomarkers

Two different types of biomarker can be useful to select patients for adjuvant chemotherapy. The first type is commonly referred to as a 'prognostic' marker, which is used to identify patients with inherently good prognosis who may be cured with locoregional therapy alone. In the context of cancer biology, prognostic tests aim to identify invasive cancers with no or little metastatic ability and, therefore, low risk of distant metastatic relapse after surgery. The second type is the 'predictive' biomarker, which allows the identification of individuals who benefit from a particular drug or treatment regimen. In the context of adjuvant therapy, benefit means improved recurrence-free and overall survival, and this is best assessed in a randomized clinical trial. The most appropriate statistical approach to examine the predictive value of a novel biomarker is to perform a test for interaction. This test compares the relative benefit of treatment according to biomarker status. By comparing the relative benefits expressed as hazard ratios, this test does not necessarily provide information on the clinical usefulness of the predictor. For example, adjuvant chemotherapy might be associated with a hazard ratio of 0.5 for death in biomarker group A and a hazard ratio of 0.75 in group B. A test for interaction can indicate that these hazard ratios are significantly different; however, both groups clearly benefited from therapy even though group B benefited more. When considering the clinical usefulness of a biomarker, examining the relative difference and the absolute difference in outcome between the marker-positive and marker-negative groups is of equal importance.

Estrogen receptor as a predictor of benefit from adjuvant chemotherapy

Currently, the clinically most useful molecular classification of breast cancer is based on assessing ER and HER2 status. Patients with ER-positive cancers can benefit from endocrine therapy, whereas those with ER-negative cancers do not; therefore, ER status is a good predictor of response to endocrine therapy. In the absence of endocrine therapy, however, a positive ER status does not predict a substantially better prognosis and, therefore, it is not a good prognostic marker in untreated patients.⁶ Meta-analysis by the Early Breast Cancer Trialists' Collaborative Group also demonstrated greater relative benefit from adjuvant chemotherapy in ER-negative disease than in ER-positive disease.¹ In women below 50 years of age, the 5-year absolute reductions in recurrence were 13.2% in patients with ER-negative cancers and 7.6% in those ER-positive cancers. In women older than 50 years, the absolute gains from chemotherapy were smaller—9.6% for ER-negative tumors, and 4.9% for ER-positive tumors—but were significantly better than with no chemotherapy. A retrospective analysis of three adjuvant trials conducted by the Cancer and Leukemia Group B and the US Intergroup (CALGB 8541, 9344, and 9741) included over 6,000 patients; the hazard ratio reduction was twice as high in women with ER-negative cancers than in those with ER-positive cancers.¹⁰ This study also showed that the difference in hazard ratio reduction was independent of the chemotherapeutic agent. Results from preoperative chemotherapy trials confirm that ER-negative breast cancers are generally more sensitive to cytotoxic therapy. A recent review of over 1,000 women who received preoperative chemotherapy showed substantial differences in pathologic complete response rates between different regimens, but it also indicated on average 2–4-fold higher pathologic complete response rates in ER-negative compared with ER-positive breast cancer for any particular regimen.¹¹ Similar results were reported by the National Surgical Adjuvant Breast and Bowel Project (NSABP) B27 clinical trial.¹² Although not all adjuvant trials reported greater benefit in the ER-negative patients, in aggregate the clinical observations are remarkably consistent with the emerging molecular data, which indicates that ER-negative and ER-positive breast cancers are fundamentally different diseases with distinct chemotherapy sensitivity.^{13, 14} ER status is only moderately useful in clinical decision-making when adjuvant chemotherapy is considered, however, because both women with ER-positive and those with ER-negative breast cancers can benefit from adjuvant chemotherapy, although to a different extent. In order to estimate the absolute benefit from chemotherapy for ER-positive breast cancer, other clinical variables including histological grade, nodal status and tumor size must also be considered. Furthermore, knowledge of ER status may give some information on general chemotherapy sensitivity, but it does not help in selecting one regimen over another.

Current ER measurements have several limitations, despite their clinical value in selecting patients for endocrine therapy and identifying subsets of patients who may be more or less sensitive to chemotherapy. Unfortunately, the existing immunohistochemistry assays show substantial variation both within and between laboratories.¹⁵ Furthermore, ER expression levels have only modest positive predictive value (30–60%) regarding response to single-agent hormonal therapies.^{16, 17} ER status provides no information about what particular chemotherapy regimen should be selected for adjuvant therapy. These limitations have driven efforts to develop more accurate and more reliable predictors of benefit from hormonal therapy, and to better define the subset of ER-positive patients who require adjuvant chemotherapy. More accurate quantification of ER expression may represent an improvement over the semiquantitative ER immunohistochemistry for assessing response. Quantitative reverse transcription–polymerase chain reaction (RT-PCR) and DNA microarrays measure ER mRNA expression over a broad dynamic range. Since ER protein and ER mRNA levels closely correlate, it is reasonable to assume that mRNA-based ER status determination will be of clinical value.¹⁸

The mere presence of ER does not guarantee that the receptor is functionally active, and other molecular events, unrelated to ER signaling, could also influence sensitivity to hormonal therapy. The Oncotype DX^® (Genomic Health. Inc., Redwood City, CA) RT-PCR-based assay represents an important diagnostic advance for ER-positive breast cancers and is also the first multigene diagnostic assay that has become commercially available in the US.¹⁹ This assay measures not only ER mRNA expression in a highly quantitative and reproducible manner but also the expression of several downstream ER-regulated genes (e.g. PR, BCL2, and SCUBE2), which could contain information on ER functionality. The same assay also quantifies HER2 expression and the expression of several other genes involved in proliferation and invasion. Combining information from each of these measurements into a single prediction score provides a method of outcome prediction that is superior to analysis of ER levels by immunohistochemistry alone. A study examined the correlation between the Oncotype DX^® Recurrence Score^® (Genomic Health, Inc.) and the likelihood of distant relapse after 5 years of tamoxifen therapy in ER-positive, node-negative, tamoxifen-treated patients who were enrolled in the NSABP randomized clinical trial B14.¹⁹ The recurrence score was predictive of relapse after tamoxifen treatment as well as overall survival, and this was independent of patient age and tumor size. These results were confirmed in a similarly treated community-based patient population in a separate study.²⁰ A recent report examined the value of the recurrence score for predicting benefit from adjuvant CMF chemotherapy in 651 patients with ER-positive, node-negative breast cancer included in the NSABP B20 randomized study.²¹ The investigators reported a significant interaction between higher recurrence score and greater benefit from adjuvant CMF chemotherapy (test for interaction: P = 0.038). The hazard ratio for distant recurrence after CMF chemotherapy was 1.31 (95% CI 0.46–3.78) for patients with a recurrence score below 18, and 0.26 (95% CI 0.13–0.53) for patients with a recurrence score of more than 31. The absolute improvement in 10-year distant-recurrence-free survival was 28% (60% versus 88%) in patients with a recurrence score above 31, while there was no benefit in patients with a recurrence score below 18. Interestingly, in this study, neither the quantity of ER measured in the ligand-binding assay nor tumor grade were predictive of benefit from chemotherapy. These results suggest that the recurrence score could identify a subset of women with ER-positive and node-negative breast cancer who have high risk of recurrence with tamoxifen therapy alone, independent of grade and quantitative ER status, and this risk can be reduced with administration of adjuvant CMF chemotherapy.

Several other multigene prediction scores are currently being explored to refine the predictive value of ER status and assess residual risk of relapse after adjuvant hormonal therapy. These tests include assessment of a two-gene expression ratio (HOXB13:IL17RB), a 44-gene signature, a 200-gene ER reporter index, and a 97-gene genomic grade index.^{22, 23, 24, 25} Each of these genomic tests has shown some ability to risk stratify ER-positive patients who received tamoxifen therapy; none, however, has undergone the same degree of clinical validation as the Oncotype DX^® assay. These tests probably represent a nascent generation of diagnostic tools that will eventually complement existing clinicopathologic variable-based prediction models.

Unsupervised classification of breast cancer

Hierarchical clustering, multidimensional scaling and self-organizing maps are commonly used mathematical tools applied to analysis of gene expression data sets to identify groups of cases that exhibit similar molecular features.²⁶ Soon after application of the first transcriptional profiling techniques it became obvious that expression of most genes in breast cancer show relatively modest variation across specimens and therefore contribute little to defining new molecular subsets. A relatively small subset of genes showing varied expression can capture most of the variation between cases.

The first study to examine comprehensive gene expression patterns in breast cancer used hierarchical clustering focusing on a set of 1,753 highly variably expressed genes.²⁷ This study included 3 normal breast samples and 40 different breast tumors including 20 repeated measurements from the same tumor. The investigators observed two major clusters in the data with additional smaller clusters within the two main clusters of the dendrogram. They suggested a four-way classification of breast cancer: luminal-like (expressing luminal cytokeratins 8 and 18), basal-like (characterized by cytokeratins 5 and 17), HER-positive (mostly, but not all, HER2 amplified) and normal-like. Subsequently, the gene list was modified and 476 genes, termed the 'intrinsic gene set', were used in a follow-up study, which included 78 cancers that suggested three additional subgroups within the luminal category: luminal-A, luminal-B, and luminal-C.²⁸ In a third study, the same investigators used a further modified version of the intrinsic gene set that included 534 genes, which was tested on 115 cases. The basal-like, normal-like, HER2-positive and two categories of luminal cancers were observed again.²⁹ A more recent publication indicates yet another gene list that is now extended to include 1,300 genes that may be applied for hierarchical clustering across multiple platforms.³⁰

Studies by others also confirmed that there are large-scale gene expression differences between ER-positive and ER-negative breast cancers and suggested that further molecular subsets may also exist within or in addition to these categories.^{31, 32} The prognosis and chemotherapy sensitivity associated with the different molecular subgroups also appears to be different. Luminal-type cancers tend to have the most favorable long-term survival, whereas the basal-like and HER2-positive tumors are most sensitive to chemotherapy but have the worst prognosis overall.^{28, 32, 33}

These studies offer an exciting new avenue for translational research with potentially important clinical implications; however, the prognostic value of classifying breast cancers based on their molecular subtypes (e.g. luminal, basal-like and other categories) has not been fully determined. This classification schema, however, provides a very important new framework for the study of breast cancer. It is no longer appropriate to consider breast cancer as a single disease with heterogeneous ER and HER2 expression. More appropriately, this form of cancer represents at least three molecularly and clinically clearly distinct diseases that perhaps arise from different precursor cells in the breast. It is critically important to develop standard methods to define molecular class and assign a molecular category to new cases. This standardization of methodology and definition of further reproducible molecular subsets within the three larger categories is yet to be accomplished.

Several limitations of the current molecular classification methods must be considered. The true prognostic or predictive value of the various molecular classes is unknown because there is a strong correlation between molecular class and conventional histopathologic variables. For example, in one study, all luminal-type cancers were ER-positive and 63% of these were also low or intermediate grade, in contrast to 95% of basal-like cancers that were ER-negative, 91% of which were high grade.³² These associations partly explain the different clinical outcome observed in different molecular classes. Furthermore, all of the above studies included patients with mixed treatment histories—a few treated with surgery alone, others with adjuvant hormonal therapy, chemotherapy or the combination of both. Larger studies including untreated patients will be needed to determine the true independent prognostic value of molecular class. In order to perform these studies, consensus must be reached on the number of molecular classes that exist and how to define them. Currently, the number of molecular subtypes of breast cancer based on mRNA expression profiles is not certain, and there is no standardized method for assigning molecular class to a new case.

Hierarchical clustering algorithms group cases together based on gene expression similarities, and the results are displayed as a dendrogram. Clustering algorithms always detect clusters, however, even if the gene expression data is random, and therefore it is critical to apply some statistical tests to determine how many robust and reproducible clusters are in a particular dendrogram. Analysis of the result with the naked eye is an appealing but less than optimal way to determine the number of clusters. There are several statistical methods that can be used to estimate the number of robust clusters in microarray data.^{34, 35, 36} Unfortunately, these statistical methods were rarely applied in earlier microarray publications. Figure 1 illustrates the inherently unstable nature of hierarchical clustering results. Figure 1A shows hierarchical clustering of 82 breast cancers using the original 689 highly variably expressed 'intrinsic genes', and Figure 1B shows the results of five different statistical methods for assessing the optimal number of clusters in the data. Each method suggests that the number of robust clusters is no more than two to three. The gene set used for clustering also affects the dendrogram results. Figure 1C illustrates how a dendrogram changes if a different set of highly variable genes (using the most recent 1,300-gene set³⁰) are used for clustering of the same data, and shows that many cases, which previously clustered together (when the original 689 genes were used), are now dispersed into other clusters. A further problem is that when new cases are added to an existing data set, the previous order of clustering is revised and a completely new dendrogram is generated even when the same genes are used in the clustering analysis. This change occurs because a clustering algorithm works by linking the two most similar specimens together and then successively merging other specimens in order of similarity. Figure 1D illustrates how the dendrogram shown in Figure 1A changed after 51 new cases were added to the data; the same 689-gene set was used for clustering as in Figure 1A. The HER2-positive, luminal, and basal-like groups remain distinguishable, but the previously normal-like group is no longer apparent. Reorganization of cases within the sub-branches of the three main clusters is even more dramatic, suggesting that, with a relatively small sample size, subtle subdivision of a dendrogram tree into luminal-A or luminal-B subtypes is not stable. Undoubtedly, there are transcriptional differences that characterize distinct molecular subgroups within the luminal, basal-like and HER2-positive groups; however, a substantially larger sample size is required to reliably identify these differences and define smaller molecular subsets of breast cancer.

Figure 1 Molecular classification of breast cancer through hierarchical clustering with "intrinsic genes".

(A) Dendrogram of 82 cases using 689 intrinsic genes. The four major clusters are color coded and represent the HER2-positive (brown), luminal (blue), normal-like (green) and basal-like (orange) molecular classes.²⁸ (B) Statistical analysis of the dendrogram result from panel A with five different methods suggests the presence of no more than two to three robust clusters. This is indicated by the sudden change in the significance scores assigned to any clusters more than two in a given data set by all statistical methods.^{34, 35, 36} (C) Hierarchical clustering of the same 82 cases with 1,300 'revised intrinsic genes' shows substantial changes to the original dendrogram. Cases are color coded by the original molecular class as shown in (A). (D) Hierarchical clustering of 133 cases using the 689 intrinsic genes including the original 82 color-coded cases and 51 new cases. Addition of new cases leads to substantial reorganization of the dendrogram. Permission obtained from the American Association for Cancer Research Inc. © Rouzier R et al. (2005) Clin Cancer Res 11: 5678–5685.

Full figure and legend (54K)Figures & Tables indexDownload Power Point slide (260K)

It is also important to realize that hierarchical clustering is not a method to use to assign a molecular class to a new case; rather, this method is one of the many tools for defining which molecular classes exist within a particular data set. After the molecular classes are agreed upon, the gene expression differences that define these groups can be used to develop a molecular class predictor that could prospectively assign the molecular class to new cases. Only after a standard method for class prediction has been developed, which includes the gene expression profiling platform, the data normalization process, the gene set, and the prediction rules, can the value of molecular classification be tested appropriately in clinical trials. Important efforts were recently made to develop such single sample class predictors.^{29, 30}

The complex and technology intense nature of molecular classification using transcriptional profiling resulted in numerous efforts being made to develop immunohistochemical markers for molecular classification. One approach simply assigns new names to previously recognized subsets of breast cancer. ER-positive cancers become luminal cancers and the ER-negative, progesterone receptor-negative, and HER2-negative (often referred to as 'triple negative') tumors are equated with basal-like cancers. This simple schema has some value because it draws attention to the fundamentally different nature of ER-positive and ER-negative cancers but probably represents oversimplification of the molecular complexity. Combination of various immunohistochemical markers including cytokeratins, epidermal growth factor receptor, ER status and HER2 status have also been proposed to define luminal, basal-like and other subtypes of breast cancer.^{37, 38, 39} Although the terminology has been adopted from gene expression studies, it is important to recognize that the molecular classes defined by immunohistochemistry correspond only partially to the molecular classes defined by gene expression profiles. In summary, various partially overlapping molecular classification methods currently exist that are poorly standardized. The availability of different methods could lead to confusing results in the future, because different investigators might assign a different class to the same cancer depending on the method used.

Supervised predictors of prognosis and response to chemotherapy

Supervised class predictors are sometimes referred to as prognostic or predictive gene signatures. Development of these signatures begins by defining a patient subset according to a particular outcome (e.g. complete response to a particular preoperative chemotherapy), and the gene expression profiles of this group are compared with those of the rest of the cases to identify differentially expressed genes using some variation of two-sample t-test or similar methods. The probability of false discovery associated with particular P values derived from the statistical tests can be estimated, and differentially expressed genes with small false discovery rates are used in combination with class prediction algorithms to formulate prediction rules based on the available training data.⁴⁰ This approach to developing a predictor could be applied to any clinical outcome of interest and could include all breast cancers or only selected subsets defined by ER status, molecular class, or any other parameter. Whether it is better to develop a clinical predictor from all samples or only from particular subsets of breast cancer depends on whether the predictor is likely to work equally well in each subset or not. It is important to consider that a particular treatment could work equally well in different molecular subsets of cancer. For example, HER2 amplification predicts response to trastuzumab in both ER-positive and ER-negative breast cancers. When the discovery sample size is limited, the most expeditious use of the samples may be to include all cases and adjust for group status during the prediction model building. There are examples of successful application of both approaches, however, in the pharmacogenomic literature. A 70-gene prognostic predictor was developed from 78 breast cancers that were selected to include an even number of cases who relapsed in 5 years or not, but neither ER status nor other clinical variables were considered when the molecular predictor was developed.⁴¹ This general prognostic predictor worked well when validated in independent cases.^{42, 43} Other investigators took a different approach and identified genes that were associated with relapse separately for the ER-negative and the ER-positive subsets of patients. The markers that were selected from each group were then combined to form a single 76-gene prognostic signature. This predictor also performed well when tested on independent cases.^{44, 45}

Supervised analyses of gene expression data were also used to explore whether chemotherapy response predictors can be developed. Neoadjuvant (preoperative) chemotherapy represents an attractive clinical setting in which to develop predictors of response to therapy. Several small studies provided 'proof of principle' that the gene expression profiles of chemotherapy-sensitive cancers are different from resistant tumors. In one study, investigators analyzed the gene expression signatures from core needle biopsies taken from 24 patients with locally advanced breast cancer prior to treatment and assessed tumor response to docetaxel. The investigators identified 92 genes that significantly correlated with the volume of residual cancer, and a genomic predictor was developed.⁴⁶ Another multigene signature of docetaxel sensitivity was also reported.⁴⁷ In another study, gene expression signatures from fine needle biopsies of 82 breast cancers were used to develop a multigene predictor of pathologic complete response to preoperative sequential weekly paclitaxel followed by 5-fluorouracil, doxorubicin, and cyclophosphamide (FAC) chemotherapy, which was validated on an independent set of 52 cases.^{48, 49} Paraffin-embedded biopsy materials from 89 patients were also analyzed by RT-PCR to identify which of 384 candidate genes were associated with a complete pathologic response. Univariate analysis revealed 86 genes that were associated with response to sequential doxorubicin and paclitaxel-based chemotherapy.⁵⁰ Multigene predictors of response to single-agent paclitaxel, doxorubicin plus cyclophosphamide, and epirubicin plus cyclophosphamide preoperative chemotherapies were also reported.^{51, 52, 53} These results represent exciting but preliminary data indicating that gene signatures might identify patients who are sensitive to a given chemotherapy regimen. The discovery and validation of chemotherapy response predictors have been slower than those of the prognostic signatures because of the limited number of archived tissues available for such research. Existing frozen tumor banks rarely contain tissues from patients who received preoperative chemotherapy with a uniform regimen; therefore, predictor studies require prospective collection of biopsies.

Molecular subclasses and markers of resistance to chemotherapy

An interesting hypothesis in predictive marker research is that different predictors might be optimal for different molecular subsets of breast cancer.³² It is hypothesized that different molecular mechanisms are responsible for sensitivity or resistance to chemotherapy in the different molecular subsets of cancers. If this is the case, examination of the overall population for marker and treatment interaction could weaken the power of a discovery study. An example of such a case includes clinical correlative studies yielding conflicting results about the value of topoisomerase II alpha (TOP2A) or HER2 amplification as predictors of response to anthracyclines.⁵⁴ Interestingly, one study noted that coamplification of these two genes was associated with the most benefit from anthracycline-based chemotherapy, whereas HER2 or TOP2A amplification alone was not significant.⁵⁵ This evidence suggests that the predictive value of TOP2A might be the most relevant in the HER2-amplified cases. Recent data from a randomized trial are consistent with this hypothesis.⁵⁶ Another possible example of a molecular-subset-specific predictive effect involves p53. Mutations in p53 and BRCA1 (both somatic and germ-line) are more common in basal-like breast cancers than in other subtypes, and these cancers are also more sensitive to anthracyclines.³² Laboratory studies suggest that normal p53 function is necessary for initiation of cell death in BRCA1 wild-type cell lines, whereas cell death is p53-independent in BRCA1-mutated cell lines.⁵⁷ Based on these in vitro observations, one could hypothesize that p53 mutations predict resistance in luminal cancers, whereas such an association might not be present in basal-like cancers. Indeed, several reports showed that mutated p53 correlated with resistance to anthracyclines;⁵⁸ however, another study suggested that mutated p53 correlated closely with the efficacy of anthracycline-based chemotherapy.⁵⁹ One possible explanation for the discrepant results is that the first study mostly included luminal cancers (in which p53 correlates with resistance) and the second study inadvertently included more basal-like cancers and, therefore, p53 mutation status simply served as a surrogate to identify the generally chemotherapy-sensitive molecular subtype. Remarkably, comparison of the patient populations of the two studies showed substantial differences—72% of the patients included in the Geisler et al. study⁵⁸ had low-grade or intermediate-grade tumors, whereas in the Bertheau et al. study⁵⁹ most patients were ER-negative and high-grade. These observations suggest that testing single bio-markers in predefined molecular subsets of breast cancer or applying supervised outcome prediction methods to specific subsets could identify molecular-class-specific predictive signatures.

Are pharmacogenomic tests better than conventional clinical parameters?

The Oncotype DX^® Recurrence Score^®, molecular subclasses, and multigene chemotherapy response predictors are all correlated with some clinical characteristics, mainly ER status and tumor grade. It is customary to report results from multivariate analyses that contain clinical parameters, and such analyses usually indicate at least some independent predictive value for the new molecular assays. Similar to predictive genes, clinicopathologic variables (e.g. as measured by the Nottingham Prognostic Index, Adjuvant! Online, and the pathologic complete response predictor nomogram) also provide better correlations when used in combination with one another.⁹ Currently, no study has compared directly the accuracy of pharmacogenomic predictors and state-of-the-art multivariable clinical prognostic and prediction models. Perhaps a more important challenge is to find out how to combine these new tools with routinely available clinical information to optimize medical decision-making. It might be useful to consider the development of various emerging genomic tests as a process that is conceptually similar to therapeutic drug evaluation (Figure 2). In this context, the goal of a phase I pharmacogenomic discovery study is to define the predictive gene set, establish the prediction rules and determine the assay cutoff points in a well-defined patient population. During phase II, the performance of the a priori defined predictor is assessed on independent cases, and the reproducibility and robustness of the assay is examined. These initial phases of marker development could be performed on archived specimens with a prospectively defined analysis plan. The phase III marker evaluation study ideally would include a prospectively conducted randomized clinical trial to show that the use of assay results yields better clinical outcome than current clinical decision-making without molecular assay results.

Figure 2 Overview of the clinical development stages of various genomic predictors.

Abbreviations: AC, doxorubicin and cyclophosphamide; EC, epirubicin and cyclophosphamide; ER, estrogen receptor; FAC, 5-fluorouracil, doxorubicin and cyclophosphamide.

Full figure and legend (28K)Figures & Tables indexDownload Power Point slide (232K)

Two clinical trials are underway that aim to determine the clinical value of using the Oncotype DX^® recurrence score and the MammaPrint^® (Agendia, Amsterdam, The Netherlands) prognostic signature in medical decision-making, and therefore could be considered phase III marker validation trials. The North American TAILORx study will randomize patients with intermediate recurrence score values to receive hormonal therapy alone or hormonal therapy plus chemotherapy to find out whether adjuvant chemotherapy improves survival in this subset of patients. In this trial, patients with low predicted risk of relapse after tamoxifen therapy will receive endocrine therapy alone, and those with high predicted risk of relapse will be treated with chemotherapy plus endocrine therapy. This trial will be part of a new National Cancer Institute initiative, the Program for the Assessment of Clinical Cancer Tests (PACCT), which seeks to facilitate individualized cancer treatment by evaluating emerging predictive and prognostic tests. The European MINDACT (Microarray In Node negative Disease may Avoid ChemoTherapy) trial will directly compare a gene-signature-based decision with an Adjuvant!-Online-based decision for cases in which the two models yield different prognostic prediction results. Both of these trials examine prognostic predictors and leave the question open as to whether one particular chemotherapy regimen might be more beneficial for some individuals than for others. As described previously within this article, the chemotherapy response prediction field is several years behind the prognostic predictors, and no predictive signature is validated sufficiently to use for prospective treatment selection (phase III evaluation). Nonetheless, several predictive signatures have been proposed and at least one of them, a paclitaxel–FAC predictor, is currently being evaluated in an independent validation study at the MD Anderson Cancer Center.